FlexRouter: Learning Complementary Model Sets for Flexible LLM Routing
FlexRouter selects model groups for coverage, not just the individually highest-scoring LLMs.
The paper argues that standard routers can waste calls on models with overlapping failure modes. FlexRouter uses determinantal point processes to balance model competence against redundancy, then greedily adds models based on marginal log-determinant gains. Its objective is the chance that at least one selected model answers correctly, matching pipelines with multiple candidates and a verifier or user downstream. The authors report higher coverage and lower redundancy than strong baselines on RouterEval, including out-of-domain tasks. HF Daily Papers' note
The paper argues that standard routers can waste calls on models with overlapping failure modes. FlexRouter uses determinantal point processes to balance model competence against redundancy, then greedily adds models based on marginal log-determinant gains. Its objective is the chance that at least one selected model answers correctly, matching pipelines with multiple candidates and a verifier or user downstream. The authors report higher coverage and lower redundancy than strong baselines on RouterEval, including out-of-domain tasks. HF Daily Papers' note
score 4